Integration of the human lung cell atlas with metacells

title: “Integration of the human lung cell atlas with metacells” author: “Leonard Herault, Aurelie Gabriel, Mariia Bilous, David Gfeller” date: “2023-11-17” output: md_document: toc: true # table of content true — =====================

In this example we will work with the Human Cell Lung Atlas core HLCA gathering around 580,000 cells from 107 individuals distributed in 166 samples.

The aim of this tutorial is to show how you can use metacells to analyze a very large dataset using a reasonable amount of time and memory. For this we will use here SuperCell via the MCAT command line tool.

Be sure to be in the MetacellAnalysisToolkit environment when you are running this Rmarkdown.

Setting up the environment

library(Seurat)

## The legacy packages maptools, rgdal, and rgeos, underpinning this package
## will retire shortly. Please refer to R-spatial evolution reports on
## https://r-spatial.org/r/2023/05/15/evolution4.html for details.
## This package is now running under evolution status 0

## Attaching SeuratObject

library(anndata)
library(SuperCell)
library(ggplot2)

color.celltypes  <- c('#E5D2DD', '#53A85F', '#F1BB72', '#F3B1A0', '#D6E7A3', '#57C3F3', '#476D87',
                      '#E95C59', '#E59CC4', '#AB3282', '#23452F', '#BD956A', '#8C549C', '#585658',
                      '#9FA3A8', '#E0D4CA', '#5F3D69', '#58A4C3', "#b20000",'#E4C755', '#F7F398',
                      '#AA9A59', '#E63863', '#E39A35', '#C1E6F3', '#6778AE', '#91D0BE', '#B53E2B',
                      '#712820', '#DCC1DD', '#CCE0F5', '#CCC9E6', '#625D9E', '#68A180', '#3A6963',
                      '#968175')

Downloading the data

You can download the annotated data from cellxgene. Choose the .h5ad option after clicking on the download button for the core atlas (3 tissues, 584,944 cells).

You can use a bash command line of this form to download the data directly in the ./HLCA_data directory. You will have to update the link (obtained by clicking on download, .h5ad selection) as links are temporary.

Please note that this may take some time as the file is quite large (~6 GB)

#Uncomment to download the data in the ./HLCA_data/ directory after updating the link
#mkdir -p ./HLCA_data
#curl -o ./HLCA_data/local.h5ad "https://corpora-data-prod.s3.amazonaws.com/7bcad396-49c3-40d9-80c1-16d74e7b88bd/local.h5ad?AWSAccessKeyId=ASIATLYQ5N5XZ2V3CYXW&Signature=CI8hgXdSO2ewDXpP%2FCb7ouxW6R8%3D&x-amz-security-token=IQoJb3JpZ2luX2VjEC0aCXVzLXdlc3QtMiJGMEQCIGoOTAGVxanApGEIeRVOL%2BRK7silMZiTtgLE%2BXguyjPjAiARoOLhXmQwzwHgme2Ll0OIZK0VIrBLaH3bSbFzRzBfuSrrAwh2EAEaDDIzMTQyNjg0NjU3NSIMbCRmBRpD%2BT0U5T8%2BKsgDcLw0fAhlIgdEjdOw%2FvUOo36uXvDClcBPXmosjNUDGVIYy67gprxvikZ%2FZHqtu%2BnodejEEIIxGJw2kv0l7dcjmGgP9IFLP6WBmsGekfI7kFCkFypmZtKXqggx9stp2K3MZCrsfcEcWttsV62c690lzdiQ4UI4lUqGqXq8C7Ah1RnxfXPQJsa3YKmHs39c3mX%2BHG5Nv4rydgzhkWE7qTkGxZvqV1cLuPMz2X78zBq5GXY0HTaGvGMgAzE5OcKbqF50sxmh0pE7PGmvz1wLYN8LB6YpMbD8qCXMdP7e4uBk2yjkK23m5m%2FrMVrCWEarSh5QqrzDR347XTg%2BkVDY301ygqy3GpCTq342sTKmUZH0PRhkliGyKvakNQU4QBy6meSQORvRX1WEhn0cRYPygyD9ugK2sDqtBl0JXUlEfqSDmE%2BXGDoRFGnKiTDSvnHhVgj64h4eTUcutZFdTILwMaYGEIl1ItElCptqvYS3rmrzdvAr5nSjx%2BnK9tKt6linyh%2Bau7zc6IfQSTzZoMut%2Fw1fOuCQ%2BQmxCaEyBXzfTTrx4%2FuxyiYAkPN0vLTtSvtuklZH7O1axMTQIonnFDsnKeVnUzl3ZEgdUbxhMLL20qoGOqYBdtJOXqTiQUDX4ZH0ReubHpog%2BorDorDJ0B08Edu6k36SwuSNu6Hv8MW%2BdWFVfqs0X%2Fx74oMs8yQC8T1gSG2HrlCfLoWIBep9lA9EHq4vUBhYB4mmJ7Fsc2MdhOtof%2BzrE8b1ILxU%2Fdeliek9Aqz0uBWcfJsEu%2FlHrC1sX4P5F8nytcLxvzCTGB43mPHeqB5DZaAKC%2FY8SmSa9CJ1Njfz8n%2FIuTLv8w%3D%3D&Expires=1700662555"

First we need to specify that we will work with the MetacellAnalysisToolkit conda environment (needed for anndata relying on reticulate and the MCAT tool).

library(reticulate)
conda_env <-  conda_list()[reticulate::conda_list()$name == "MetacellAnalysisToolkit","python"]

use_condaenv(conda_env)

Splitting atlas by datasets

First we will use anndata to read in backed mode (saving a lot of memory) the whole atlas and write one h5ad file for each dataset. This should take less than 10 minutes.

If you are limited in time feel free to process only a subset of the dataset.

adata <- read_h5ad("./HLCA_data/local.h5ad",backed = "r")
adata$var_names <- adata$var$feature_name # We will use gene short name for downstream analyses
datasets <- unique(adata$obs$dat)

# If you are limited in time you can process on half of the datasets (uncomment th following line)
# datasets <- datasets[1:7]


print(dim(adata))

lapply(datasets,FUN =  function(x) {
  dir.create(paste0("./HLCA_data/datasets/",x),recursive = T)
  adata.dataset <- AnnData(X = adata[adata$obs$dataset == x]$raw$X,
                           var = adata[adata$obs$dataset == x]$var,
                           obs = adata[adata$obs$dataset == x]$obs)
  #This will allow us to construct supervised metacell for each cell type in each sample later in the second example
  adata.dataset$obs$ann <- as.character(adata.dataset$obs$ann_level_3)
  # For cell without an annotation at the 3rd level we will use the second level of annotation
  adata.dataset$obs$ann[adata.dataset$obs$ann_level_3 == 'None'] = as.character(adata.dataset$obs$ann_level_2[adata.dataset$obs$ann_level_3 == 'None'])
  adata.dataset$obs$ann_sample <- paste0(adata.dataset$obs$ann,"_",adata.dataset$obs$sample)
  
  write_h5ad(adata.dataset,paste0("./HLCA_data/datasets/",x,"/sc_adata.h5ad"))
}
)

remove(adata)
gc()

Building metacell

We build metacells with the MCAT command line using SuperCell (-t SuperCell). To facilitate downstream analysis of the donors we build metacells for each sample in each dataset (-a sample). Here we will use 2000 highly variable genes (-f 2000) to compute the PCA from which we used 50 principal components (-m 50) to build a k = 30 (-k 30) nearest neighbor graph on which the metacells are identified using a graining level of 50 (-g 50). We use an adata .h5ad output format (-s adata) as it is faster to write and lighter to store than a Seurat .rds object.

This step takes around 20 min with multiple cores (-l 6). Be aware that parallel processing requires more memory (32 GB of memory required for 6 cores).

If you are limited in memory you should still be able to process the samples by reducing the number of cores (e.g. -l 3) or by sequentially processing the samples (just remove the -l) in a slightly longer time.

for d in ./HLCA_data/datasets/*;
do ../cli/MCAT -t SuperCell -i $d/sc_adata.h5ad -o $d -a sample -l 6 -n 50 -f 2000 -k 30 -g 50 -s adata
done

Loading metacell objects

We load the .h5ad objects and directly convert them in Seurat objects to benefit from all the functions of this framework.

metacell.files <- sapply(datasets, FUN = function(x){paste0("./HLCA_data/datasets/",x,"/mc_adata.h5ad")})

metacell.objs <- lapply(X = metacell.files, function(X){
  adata <- read_h5ad(X)
  countMatrix <- Matrix::t(adata$X)
  colnames(countMatrix) <- adata$obs_names
  rownames(countMatrix) <- adata$var_names
  sobj <- Seurat::CreateSeuratObject(counts = countMatrix,meta.data = adata$obs)
  sobj <- RenameCells(sobj, add.cell.id = unique(sobj$sample)) # we give unique name to metacells
  return(sobj)
})

Merging objects and basic quality control

Given the single-cell metadata, the MCAT tool automatically assign annotations to metacells and computes purities for all the categorical variables present in the metadata of the input single-cell object.

Thus, we can check the purity of our metacells at different levels of annotations, as well as their size (number of single cells they contain).

To do so we merge the object together and use the Seurat VlnPlot function.

unintegrated.mc <- merge(metacell.objs[[1]],metacell.objs[-1])

VlnPlot(unintegrated.mc,features = c("size","ann_level_1_purity"),group.by = 'dataset',pt.size = 0.001,ncol=2)

VlnPlot(unintegrated.mc,features = c("ann_level_2_purity","ann_level_3_purity"),group.by = 'dataset',pt.size = 0.001,ncol=2)

We can also use box plots.

ggplot(unintegrated.mc@meta.data,aes(x=dataset,y=ann_level_2_purity,fill = dataset)) + geom_boxplot() +
  scale_x_discrete(guide = guide_axis(angle = 45)) 

ggplot(unintegrated.mc@meta.data,aes(x=dataset,y=ann_level_3_purity,fill = dataset)) + geom_boxplot() +
  scale_x_discrete(guide = guide_axis(angle = 45)) 

ggplot(unintegrated.mc@meta.data,aes(x=dataset,y=ann_level_4_purity,fill = dataset)) + geom_boxplot() +
  scale_x_discrete(guide = guide_axis(angle = 45)) 

ggplot(unintegrated.mc@meta.data,aes(x=dataset,y=ann_finest_level_purity,fill = dataset)) + geom_boxplot() +
  scale_x_discrete(guide = guide_axis(angle = 45)) 

Overall metacells from the different datasets present a good purity until the third level of annotation.

Unintegrated analysis

Let’s first do a standard dimensionality reduction without batch correction.

DefaultAssay(unintegrated.mc) <- "RNA"
unintegrated.mc <- NormalizeData(unintegrated.mc)
unintegrated.mc <- FindVariableFeatures(unintegrated.mc)
unintegrated.mc <- ScaleData(unintegrated.mc)
unintegrated.mc <- RunPCA(unintegrated.mc)
unintegrated.mc <- RunUMAP(unintegrated.mc,dims = 1:30)

umap.unintegrated.datasets <- DimPlot(unintegrated.mc,reduction = "umap",group.by = "dataset") + NoLegend() + ggtitle("unintegrated datasets")
umap.unintegrated.types <- DimPlot(unintegrated.mc,reduction = "umap",group.by = "ann_level_2",label = T,repel = T,cols = color.celltypes)+ NoLegend() + ggtitle("unintegrated cell types")

umap.unintegrated.datasets + umap.unintegrated.types

remove(unintegrated.mc) # we won't use the unintegrated object anymore
gc()

You can see on the plots that a batch effect is clearly present at the metacell level with metacells clustering by datasets inside the major cell types. Let’s correct it.

Seurat integration

Here we will use the standard Seurat_v4 batch correction workflow. As in the original study, we use the dataset rather than the donor as the batch parameter. See method section “Data integration benchmarking” of the original study for more details.

This should take less than 5 minutes.

# normalize each dataset 
metacell.objs <- lapply(X = metacell.objs, FUN = function(x) {
  DefaultAssay(x) <- "RNA";
  x <- RenameCells(x, add.cell.id = unique(x$sample)) # we give unique name to metacells
  x <- NormalizeData(x)
  x <- FindVariableFeatures(x, selection.method = "vst", nfeatures = 2000)
  return(x)})

features <- SelectIntegrationFeatures(object.list = metacell.objs)

metacell.objs <- lapply(X = metacell.objs, FUN = function(x) {
    x <- ScaleData(x, features = features, verbose = FALSE)
    x <- RunPCA(x, features = features, verbose = FALSE)
})

anchors <- FindIntegrationAnchors(object.list = metacell.objs, 
                                       anchor.features = features, 
                                       reduction = "rpca",
                                       reference = c(1,2,5,9,11), # the 5 biggest datasets (in term of metacell number) are used as reference
                                       dims = 1:30)

remove(metacell.objs) # We don't need the object list anymore
gc()

combined.mc <- IntegrateData(anchorset = anchors,k.weight = 50) # we have to update the k.weight parameters because the smallest dataset contain less than 100 metacells

Check the obtained object.

combined.mc

## An object of class Seurat 
## 30024 features across 11706 samples within 2 assays 
## Active assay: integrated (2000 features, 2000 variable features)
##  1 other assay present: RNA

We can verify that the sum of metacell sizes correspond to the original number of single-cells.

sum(combined.mc$size)

## [1] 584944

Seurat returns the slot "integrated" that we can use for the downstream analysis.

DefaultAssay(combined.mc) = "integrated"
combined.mc <- ScaleData(combined.mc, verbose = FALSE)
combined.mc <- RunPCA(combined.mc, npcs = 30, verbose = FALSE)
combined.mc <- RunUMAP(combined.mc, reduction = "pca", dims = 1:30)

## 12:22:40 UMAP embedding parameters a = 0.9922 b = 1.112

## 12:22:40 Read 11706 rows and found 30 numeric columns

## 12:22:40 Using Annoy for neighbor search, n_neighbors = 30

## 12:22:40 Building Annoy index with metric = cosine, n_trees = 50

## 0%   10   20   30   40   50   60   70   80   90   100%

## [----|----|----|----|----|----|----|----|----|----|

## **************************************************|
## 12:22:42 Writing NN index file to temp file /tmp/35282353/RtmpWZTMG3/file26b8e33465da10
## 12:22:42 Searching Annoy index using 1 thread, search_k = 3000
## 12:22:45 Annoy recall = 100%
## 12:22:45 Commencing smooth kNN distance calibration using 1 thread with target n_neighbors = 30
## 12:22:46 Initializing from normalized Laplacian + noise (using irlba)
## 12:22:47 Commencing optimization for 200 epochs, with 484456 positive edges
## 12:22:53 Optimization finished

combined.mc <- RunUMAP(combined.mc, dims = 1:30,reduction =  "pca",reduction.name = "umap")

## 12:22:53 UMAP embedding parameters a = 0.9922 b = 1.112
## 12:22:53 Read 11706 rows and found 30 numeric columns
## 12:22:53 Using Annoy for neighbor search, n_neighbors = 30
## 12:22:53 Building Annoy index with metric = cosine, n_trees = 50
## 0%   10   20   30   40   50   60   70   80   90   100%
## [----|----|----|----|----|----|----|----|----|----|
## **************************************************|
## 12:22:54 Writing NN index file to temp file /tmp/35282353/RtmpWZTMG3/file26b8e334f0d51d
## 12:22:54 Searching Annoy index using 1 thread, search_k = 3000
## 12:22:57 Annoy recall = 100%
## 12:22:57 Commencing smooth kNN distance calibration using 1 thread with target n_neighbors = 30
## 12:22:58 Initializing from normalized Laplacian + noise (using irlba)
## 12:23:00 Commencing optimization for 200 epochs, with 484456 positive edges
## 12:23:05 Optimization finished

Now we can make the plots and visually compare the results with the unintegrated analysis.

umap.integrated.datasets <- DimPlot(combined.mc,reduction = "umap",group.by = "dataset") + NoLegend() + ggtitle("integrated datasets")
umap.integrated.celltypes <- DimPlot(combined.mc,reduction = "umap",group.by = "ann_level_2",label = T,repel = T,cols = color.celltypes) + NoLegend() + ggtitle("integrated cell types")

umap.integrated.datasets + umap.integrated.celltypes + umap.unintegrated.datasets + umap.unintegrated.types

Seurat efficiently corrected the batch effect in the data while keeping the cell type separated, but other batch correction methods such as harmony would have also done the job.

Note that In the original study, datasets were integrated using SCANVI semi-supervised integration using partial annotation obtained for each dataset prior integration. If you are interested in such supervised approach at the metacell level in R you can have a look to our second example using the STACAS package.

We can navigate in the different annotation levels.

library(ggplot2)

DimPlot(combined.mc,group.by = "ann_level_1",reduction = "umap",label = T,repel = T,cols= color.celltypes) + NoLegend()

DimPlot(combined.mc,group.by = "ann_level_2",reduction = "umap",label = T,repel = T,cols= color.celltypes) + NoLegend()

DimPlot(combined.mc,group.by = "ann_level_3",reduction = "umap",label = T, repel = T,cols= color.celltypes) + NoLegend()

Downstream analysis

Clustering

We cluster the metacells based on the corrected PCA space by Seurat.

DefaultAssay(combined.mc) <- "integrated"
combined.mc <- FindNeighbors(combined.mc,reduction = "pca",dims = 1:30)
combined.mc <- FindClusters(combined.mc,resolution = 0.5) 
UMAPPlot(combined.mc,label=T) + NoLegend()

Differently expressed gene (DEG) analysis.

Now let’s found the markers of the cluster 19 we’ve just identified.

DefaultAssay(combined.mc) <- "RNA"
markers18 <- FindMarkers(combined.mc,ident.1 = 18,only.pos = T)

## For a more efficient implementation of the Wilcoxon Rank Sum Test,
## (default method for FindMarkers) please install the limma package
## --------------------------------------------
## install.packages('BiocManager')
## BiocManager::install('limma')
## --------------------------------------------
## After installation of limma, Seurat will automatically use the more 
## efficient implementation (no further action necessary).
## This message will be shown once per session

head(markers18)

##          p_val avg_log2FC pct.1 pct.2 p_val_adj
## TNFRSF17     0  0.9274364 0.762 0.029         0
## TCL1A        0  1.0045973 0.481 0.019         0
## CD79A        0  2.9087361 1.000 0.133         0
## FCRLA        0  0.7458042 0.801 0.018         0
## BLK          0  0.8766954 0.845 0.054         0
## FCRL5        0  0.8790523 0.934 0.020         0

This cluster clearly present a B cell signature with marker genes such as CD19 and PAX5

genes <-c("CD19","PAX5") # knwon mast cells markers 
markers18[genes,]

##              p_val avg_log2FC pct.1 pct.2     p_val_adj
## CD19 3.323011e-222  0.7754585 0.812 0.111 9.312405e-218
## PAX5  0.000000e+00  0.3968396 0.630 0.022  0.000000e+00

VlnPlot(combined.mc,genes,ncol = 1)

By looking at the metacell annotation (assigned from the original single-cell metadata by MCAT), we can verify that we correctly retrieved the B cell lineage cluster

DimPlot(combined.mc[,combined.mc$integrated_snn_res.0.5 == 18],group.by = c("ann_level_3","integrated_snn_res.0.5"),ncol = 2)

Cell type abundances analyses.

We can easily make analysis of cell type abundances for different clinical variables as we construct metacell by sample. We have to take metacell size into account for these analyses. For instance we can analyse the proportion of different epithelial cell types depending on the smoking status.

library(reshape2)
combined.mc.epith <- combined.mc[,combined.mc$ann_level_1 == "Epithelial"] 
#combined.metacells$major_type <- droplevels(combined.metacells$major_type)
smpCounts <- aggregate(combined.mc.epith$size, by=list(sample = combined.mc.epith$sample,
                                                        major_type = combined.mc.epith$ann_level_3,
                                                        smoking_status = combined.mc.epith$smoking_status),
                                                        FUN=sum)

remove(combined.mc.epith)
gc()

ggplot(smpCounts,aes(x = smoking_status,fill=major_type)) + geom_bar(position = "fill") + scale_fill_manual(values = color.celltypes) + xlab("% epithelial cells")

Samples from smokers seem to present more AT2 cells but this quick analysis is for illustrative purposes only. In practice it’s far more complex to draw conclusion as we should have considered the variations between samples/donors as well as many other technical (tissue dissociation protocol, tissue sampling method, single-cell platform, … ) and biological (BMI, sex, Age, …) variables.

Conclusion

Overall we made a precise simplification of the original atlas using metacells built from each sample separately. By reducing the size of the original atlas by a factor of 50 we could load the data, make an integration to correct batch effect and recapitulate the main different cell types using a reasonable amount of time and memory. In contrast, simply loading the original single-cell data in R using Seurat is extremely time-consuming and challenging even for the most powerful computers.

In this first example we used a fully unsupervised workflow and did not use any prior biological knowledge. Authors of the original study made a remarkable work annotating the hundreds of thousands cells of the atlas. In the second example we propose a supervised workflow using this annotation to guide both the metacell identification and the batch correction.

We can save the results for comparison with the second example.

saveRDS(combined.mc,"./HLCA_data/combined.mc.unsup.rds")